EEG-DGRN: dynamic graph representation network for subject-independent ERP detection

• Published in: Brain-apparatus communication Vol. 4; no. 1
• Main Authors: Zhu, Jiabin, Jin, Xuanyu, Ming, Yuhang, Kong, Wanzeng
• Format: Journal Article
• Language: English
• Published: Taylor & Francis Group 31.12.2025
• Subjects: Brain–computer interface; dynamic graph neural network; electroencephalogram; event-related potential; subject-invariant representation
• ISSN: 2770-6710; 2770-6710
• DOI: 10.1080/27706710.2024.2447576

Objectives The inter-subject variability remains a formidable challenge in electroencephalogram (EEG) signal processing. Existing event-related potential (ERP) detection methods inadequately consider the dynamic connectivity of EEG signals and event response differences between subjects, limiting the discriminability of task-related features.Methods In this article, we propose EEG-DGRN, a dynamic graph representation network designed for subject-independent ERP detection. Specifically, the dynamic graph mechanism is used to capture the task-relevant connectivity relationship between EEG channels over time. Then, considering the local and global topology structure, a dual-branch graph pooling module is employed to prune features from different granularity. After that, the temporal dynamic attention module enables the model to pay more attention to subject-invariant representations.Results Our EEG-DGRN model is evaluated on a publicly available rapid serial visual presentation dataset. It achieves a remarkable mean balanced classification accuracy of 87.05%, outperforming all other methods compared in this study.Conclusion Such performance demonstrates its ability to extract subject-invariant EEG features and generalize effectively to unseen subjects. Lastly, ablation studies confirm the effectiveness of each module in EEG-DGRN, highlighting their contributions to the overall performance.